Audit Trails: How Much Visibility Does Muse Actually Give You?

An agent that keeps working after you close the app is, by definition, doing things you didn't watch happen. That's the entire value proposition and the entire risk in one sentence, and every company launching an always-on agent this September had to answer the same question: how does a user find out what their agent actually did while they were away? Meta's answer for Muse is an audit trail — described in the company's own announcement as showing "a complete audit trail of everything it has done and plans to do." That's a strong claim, and it's worth pulling apart into what it structurally implies versus what remains unstated.

What the audit trail is attached to

The audit trail isn't a standalone feature — it's the visible half of the Sentinel system. Sentinel is a separate agent that runs alongside Muse inside the same Secure VM, gating every outbound network request and connected-app action before it happens. Meta states that "nothing Muse does reaches the internet unless the Sentinel approves it, and it asks the person for permission when needed." If that's accurate, the audit trail is effectively a log of Sentinel's decisions — every approved action, and by extension every point where the agent asked for and received a human's go-ahead. That's a meaningfully different design than a generic activity log bolted onto an agent after the fact; it implies the audit record and the permission system are the same mechanism viewed from two directions, which is a more defensible architecture than logging actions independently of whatever approved them.

The other structural detail worth noting: users get granular, per-connection controls, not just visibility. Meta's own description gives the example of email — a person decides whether Muse can only read messages or can also send them, and can change or disconnect any connected service at any time. Combined with the audit trail, that gives a user two distinct tools: a forward-looking permission boundary (what Muse is even allowed to try) and a backward-looking record (what it actually did). That's a more complete design than either piece alone.

What "plans to do" actually means is unclear

The phrase that deserves the most scrutiny in Meta's own description is "plans to do." A trail of completed actions is a comparatively simple engineering problem — log every Sentinel-approved event with a timestamp. A trail of planned actions is a different and harder claim: it implies visibility into the agent's forward task queue, which requires the underlying model to expose its own intended next steps in a legible way, not just retroactively justify what it already executed. Meta hasn't published a screenshot, a walkthrough, or any technical detail about how far in advance the "planned" side of the trail actually looks, how it's presented, or whether it updates in real time as the agent's plan changes. Until that's shown, "plans to do" should be read as a stated design goal rather than a demonstrated capability — this is exactly the kind of undisclosed detail the launch coverage doesn't resolve.

There's also no public account of frequency or granularity. Does every read of a connected inbox log an entry, or only actions that change state? Is the trail retained indefinitely, or does it roll off after some window? None of the sources reviewed for this article — Meta's own announcement, TechCrunch's coverage of the September 23 feature rollout, or CNBC's reporting on the launch — answer these questions, which means anyone evaluating Muse for actual oversight purposes is currently taking Meta's word for the existence of a useful audit surface without being able to inspect its actual shape.

How this compares to Dots' Activity View

OpenAI's Dots ships a comparable but distinctly named feature: an Activity View for oversight and intervention, sitting alongside Custom Rules (which let a user require approval for or prohibit specific actions outright) and Auto-Review (which checks consequential actions before execution happens). That's a three-part system with named, separable functions — rules that shape what's allowed, a review step for anything consequential, and a view for after-the-fact inspection. Muse's design compresses something similar into two parts — Sentinel's real-time gate and the resulting audit trail — which is architecturally tidier but gives a user fewer named levers to configure directly. We compare both systems' overall philosophy in Muse vs Dots: Consumer Trust vs Enterprise Plumbing.

Grok Bot, by contrast, has published no equivalent oversight feature in any source reviewed for this series. xAI's own materials describe bots that "only come back when something needs your approval," which implies some form of approval gating exists, but there's no named audit trail, activity log, or after-the-fact review surface described anywhere in xAI's announcement or in independent coverage from layer3labs or netalith. That's a real, citable gap — not evidence that no such feature exists internally, but evidence that Meta and OpenAI have both chosen to make oversight a headline feature of their launches, while xAI has not, at least not yet.

The honest bottom line

Meta's audit trail claim is more architecturally credible than a marketing throwaway, because it's tied to a real enforcement mechanism (Sentinel) rather than floated as a standalone promise. But "complete audit trail of everything it has done and plans to do" is still a launch-week description, not a documented feature with published specifics on retention, granularity, or how forward-looking plans are actually surfaced. Until Meta — or an independent reviewer with hands-on access — shows what that screen actually looks like in practice, the accurate summary is: Muse has a structural reason to have a genuinely useful audit trail, and Meta says it does, and neither of those facts is the same as having seen one.

Part 25 of 30 in the Meta Muse series. Previously: The Zuckerberg Pitch: What Understands Your Goals Is Doing Rhetorically. Next: One Password and Muse: Credential Management Without the Copy-Paste. Browse more builder insights or explore AI skills at aiskill.market.